JP6430318B2 - Unauthorized voice input determination device, method and program - Google Patents

Description

  The present invention relates to a technique for determining whether or not an input voice is illegal.

  Speaker recognition by voice (hereinafter simply referred to as speaker recognition) is broadly divided into verification and identification.

  Speaker verification is used for identity verification, for example. In speaker verification, the user first declares his username to the system. Next, input voice into the system. The system determines whether the input voice is actually a user who has reported.

  On the other hand, speaker identification determines who the input voice is. In speaker identification, the person with the most similar voice is returned as a result from the pre-registered persons.

  In some cases, both verification and speaker are combined. In this case, whether or not the speaker is a non-registered speaker is determined (verified), and if it is a registered speaker, someone is specifically determined (identified). These are collectively called speaker recognition.

  There are two types of speaker recognition: text-dependent and text-independent. The text-dependent type is a format in which a user reads a predetermined sentence when performing recognition. On the other hand, the text independent type is a format in which the user may utter any word.

  In speaker recognition, voice pre-registration is required. Registration is performed for one or more utterances. For convenience, it is desirable that registration utterances are short and can be properly recognized even when the number of registrations is small.

  Speaker recognition is realized by calculating features from individual input speech and using techniques such as existing outlier detection and class classification algorithms. In the case of speaker verification, it is only necessary to determine the binary value of a registered speaker or the other, so outlier detection and binary classification algorithm can be used. Speaker identification is nothing but a multi-level classification problem. As technologies specifically used for speaker recognition, for example, technologies described in Non-Patent Documents 1 and 2 are known. Non-Patent Document 2 also describes feature quantities used for speaker recognition.

  Next, an outline of vowel section detection will be described. In the vowel section detection, a voice is input and a determination section of the vowel is detected. Processing is basically performed by paying attention to the formant component included in the speech (for example, see Non-Patent Document 3). A formant is a peak that appears in a low-frequency component when a speech is subjected to frequency analysis by Fourier transform or the like. A plurality of peaks appear and have harmonic characteristics. In general, a plurality of peaks are called first formant, second formant,... In order from the lowest frequency. The interval, size, and frequency value of these formants are generally called harmonic structures and formant structures. Formants often appear in vowels, but do not appear remarkably in consonants, which is an important feature for detecting vowel intervals.

  In fact, some consonants extract formants (voiced consonants). Therefore, the vowel section is estimated in consideration of the structure. That is, processing for determining whether or not a harmonic structure that is likely to be a vowel has been performed.

  For this reason, when a general signal, not limited to speech, is input to vowel section detection, a signal section having a harmonic structure similar to the vowel is likely to be determined as a vowel. For example, this is the sound of a musical instrument. Conversely, if there is no similar structure, the possibility of being determined as a vowel is low.

Tetsuji Ogawa and Tomoko Matsui, “Machine Learning for Speaker Recognition”, Journal of the Acoustical Society of Japan, Vol.69, No.7, pp.349-356, 2013. Wang Dragon, Masafumi Nishida, Akira Tsuji, Kanae Amino, “Robustness in Speaker Recognition”, Journal of the Acoustical Society of Japan, Vol.69, No.7, pp.357-364, 2013. Misaki Atsumi, Takayuki Arai, Nao Hodoshima, "Simple automatic detection method for vowel segments of speech-for the purpose of improving speech intelligibility in reverberant environment", Proc. -330, 2010/09.

  By the way, speaker recognition is a technique based on the premise of appropriate input of speech. Therefore, it is useful to provide the speaker recognition system with a function for determining an appropriate voice input. This is especially important during registration. This is because if the registered voice signal is inappropriate, it cannot be recognized correctly.

  The speech premised on the speaker recognition technology is a speech uttering a word, but it is difficult to define it strictly or to clearly show the point to the user for understanding. For example, it may be said that speech composed only of consonants is uttering words, but it is difficult to properly extract speaker characteristics from only consonants, and as a speaker recognition system of the current technical level, I want to regard it as invalid input. For example, “Susu sou (the vowel 'U' is not pronounced and only the consonant 's' is uttered)” corresponds to this. Furthermore, even vowels that are extremely long are technically somewhat difficult. For example, this is something that keeps saying “Ahhhh” for a few seconds. Furthermore, it is also assumed that an input in which a sound such as a sound that does not ring the throat, a sound that vibrates the tongue, a sound that sounds, a breath blowing / inhalation sound, a whistle, a lip sound, or a cough is dominant is also an illegal input. These are difficult to remove with existing speech segment detection technology.

  In the case of text-dependent speaker recognition, it is possible to determine the appropriateness of the input speech by using a technology such as speech recognition and comparing the text with the actual utterance content. However, for text-independent types, the method is not self-evident.

  An object of the present invention is to provide an unauthorized speech input determination device, an speech signal processing device, a method, and a program capable of determining whether an input speech is unauthorized even in text-independent speech signal processing. It is to be.

Fraud audio input judging apparatus according to an aspect of the present invention, the speech section detection unit for generating a speech section signal by extracting a portion of the speech interval from the speech signal inputted, vowel detecting the vowel section of the speech section signal A vowel section detection result analysis section that determines whether or not the speech signal is fraudulent based on the length of the detected vowel section , and a vowel section detection result analysis section, If the audio signal is determined not to be bad speech is transmitted to the row cormorants signal processing unit processing the talker recognition using the voice signal or the speech section signal, the input audio signal or speech section signal.
An unauthorized speech input determination device according to an aspect of the present invention includes a speech segment detection unit that generates a speech segment signal by extracting a voiced segment from an input speech signal, and a vowel that detects a vowel segment of the speech segment signal A vowel section detection result analysis section that determines whether or not the speech signal is fraudulent based on the length of the detected vowel section, and a vowel section detection result analysis section, If it is determined that the voice signal is not an illegal voice, the input voice signal or voice section signal is transmitted to a signal processing unit that performs speaker registration processing using the voice signal or voice section signal.

Audio signal processing apparatus according to one aspect of the present invention, a fraud audio input determination device described above, when the audio signal is determined not to be bad speech signal processing for processing the talker recognition using the voice signal And a section.
An audio signal processing device according to an aspect of the present invention includes the above-described unauthorized speech input determination device and a signal processing unit that performs speaker registration processing using an audio signal when it is determined that the audio signal is not unauthorized speech And.

  Even in the text-independent voice signal processing, it is possible to determine whether the input voice is illegal.

The block diagram for demonstrating the example of an unauthorized audio | voice input determination apparatus and an audio | voice signal processing apparatus. The flowchart for demonstrating the example of an unauthorized audio | voice input determination method and an audio | voice signal processing method.

[Technical background]
One feature of the present invention is that a vowel section detection technique is applied to input speech, and the application result is used to determine unauthorized speech input. In the detection of the vowel section, the property that it is difficult to determine the vowel is used unless the harmonic structure of the input sound is similar to the vowel.

  Among the illegal voice input in speaker recognition, only the consonant sound, the sound that does not ring the throat, the sound that vibrates the tongue, the sound that blows, the breath blowing / breathing sound, the lip sound, the cough, etc. Is different. Therefore, the possibility that the vowel section is detected as a vowel is small.

  In fact, speaker recognition for speech with as few vowels as possible is technically difficult, and considering a signal that does not have vowel-like features as an incorrect input in speaker recognition can be an option.

  On the other hand, in the case of illegal voice input in speaker recognition, when a vowel is extremely long or a melody is squeezed by a whistle, a continuous long section is determined as a vowel. Since normal words sandwich consonants, it is rare for long sections to be vowels. Therefore, it is possible to distinguish illegal voices from normal voices based on the length of the detected vowel section.

[Embodiment]
The audio signal processing device includes, for example, an unauthorized audio input determination device 1 and a signal processing unit 2. The unauthorized speech input determination device 1 includes, for example, a speech segment detection unit 11, a vowel segment detection unit 13, and a vowel segment detection result analysis unit 14.

<Audio section detection unit 11>
The input voice signal is passed to the voice section detection unit 11 of the unauthorized voice input determination device 1.

The voice section detection unit 11 generates a voice section signal by extracting a voiced section from the input voice signal (step S11). That is, the voice section detection unit 11 removes a silent part from the input voice signal. The generated speech segment signal is passed to the vowel segment detection unit 13.

An existing technique may be used for extracting the portion of the sound section. For example, a section having a voice signal size equal to or greater than a predetermined threshold is determined to be a voiced section, a section having a voice signal magnitude equal to or greater than a predetermined threshold is determined to be a silent section, and the determined voice A voice section signal can be generated by combining only sections.

<Vowel section detector 13>
The vowel section detector 13 detects a vowel section of the speech section signal (step S13). Vowel section information, which is information about the detected vowel section, is passed to the vowel section detection result analysis unit 14.

The vowel section information is, for example, the start time and duration of the vowel section. When a plurality of vowel sections are detected in the speech section signal, vowel section information of each of the plurality of vowel sections is generated and passed to the vowel section detection result analysis unit 14. In the following, I is an integer of 1 or more, the number of detected vowel segments is I, i = 1,..., I, and the detected duration of the i-th vowel segment is expressed as V (i). .

<Vowel segment detection result analysis unit 14>
Based on the length of the detected vowel section, the vowel section detection result analysis unit 14 determines whether the voice signal is an unauthorized voice (step S14).

  One specific configuration method of the vowel section detection result analysis unit 14 is to determine the ratio of the vowel section and determine illegal voice input by threshold processing. Now, if the length of the speech interval signal is expressed as L, the ratio R occupied by the vowel interval is

Can be calculated. The vowel section detection result analysis unit 14 determines that the input voice signal is an illegal voice when R is smaller than a predetermined threshold value TR. Otherwise, it is determined that the input audio signal is not illegal. R has a value between 0 and 1. The threshold value TR is a numerical value determined in advance by the system developer, and is set in a range that R can take. That is, the threshold value TR is set to a predetermined value between 0 and 1. If the threshold value TR is large, it is determined that many voices are illegal.

  This method is intended to detect unauthorized input that does not have a harmonic structure similar to a vowel.

  Another specific configuration method of the vowel section detection result analysis unit 14 is a method of performing threshold processing on the maximum length of the normalized vowel section. For example, the maximum value M of the length of the normalized vowel section is calculated as follows.

And calculate. The vowel section detection result analysis unit 14 determines that the voice is illegal when M is greater than a predetermined threshold value TM. Otherwise, it is determined that the input audio signal is not illegal. M has a value between 0 and 1. The threshold value TM is a numerical value determined in advance by the system developer, and is set in a range that M can take. If the threshold value TM is small, it is determined that many voices are illegal.

  This method is intended to detect illegal input such as extremely long vowels.

  The above-described method may be applied to two stages, and an audio signal that is determined to be illegal in either one may be determined to be illegal. That is, the vowel section detection result analysis unit 14 determines that the input voice signal is an illegal voice if R <TR or M> TM, and otherwise determines that the input voice signal is not an illegal voice. May be.

  Of course, unauthorized voice input may be determined using a result of vowel interval detection by a method other than the above-described method.

  When it is determined that the input voice signal is an illegal voice, the vowel section detection result analysis unit 14 notifies the user or the like, and when the input voice signal is determined not to be an illegal voice. In this case, the input audio signal or audio interval signal is passed to the signal processing unit 2.

<Signal processing unit 2>
When it is determined that the input voice signal is not an illegal voice, the signal processing unit 2 performs speaker recognition or speaker registration processing using the voice signal or the voice interval signal.

  For the speaker recognition or the speaker registration, the existing technology described in the column of background art or non-patent documents 1 and 2 may be used.

[Modification]
The unauthorized speech input determination device 1 may include a noise suppression / removal unit 12. In this case, the speech segment signal generated by the speech segment detection unit 11 is passed to the noise suppression / removal unit 12. The noise suppression / removal unit 12 suppresses or removes noise in the speech section signal (step S12). The speech segment signal from which noise has been suppressed or removed is passed to the vowel segment detection unit 13. The vowel section detection unit 13 performs processing for detecting the vowel section of the speech section signal from which noise has been suppressed or removed by the noise suppression removal unit 12 in the same manner as described above.

  An existing technique may be used for noise suppression or removal. In addition, since vowel interval detection is applied after this noise suppression or removal processing, it is possible to suppress the type of noise (for example, noise signals similar in vowel and harmonic structure) that cause a reduction in accuracy of vowel interval detection. A simple noise suppression method may be used.

  The processing described in the unauthorized speech input determination device, the speech signal processing device, and the method is not only executed in time series according to the order of description, but also in parallel or individually according to the processing capability of the device that executes the processing or as necessary. May be executed.

  Needless to say, other modifications are possible without departing from the spirit of the present invention.

[Program and recording medium]
When each processing in the unauthorized speech input determination device and the speech signal processing device is realized by a computer, processing contents of functions that each device should have are described by a program. Then, by executing this program on a computer, each process is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  Each processing means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

DESCRIPTION OF SYMBOLS 1 Fraud sound input determination apparatus 11 Voice section detection part 12 Noise suppression removal part 12 Noise suppression removal part 13 Vowel part detection part 14 Vowel part detection result analysis part 2 Signal processing part

Claims (6)

A voice section detection unit for generating a speech section signal by extracting a portion of the speech interval from the speech signal input,
A vowel section detector for detecting a vowel section of the speech section signal;
A vowel section detection result analysis unit that determines whether or not the voice signal is fraudulent based on the length of the detected vowel section,
The vowel interval detection result analysis unit, if the audio signal is determined not to be bad speech to row cormorants signal processing unit processing the talker recognition using the speech signal or the speech section signal, the Transmitting the input voice signal or the voice interval signal ,
Illegal voice input determination device. A voice section detection unit for generating a speech section signal by extracting a portion of the speech interval from the speech signal input,
A vowel section detector for detecting a vowel section of the speech section signal;
A vowel section detection result analysis unit that determines whether or not the voice signal is fraudulent based on the length of the detected vowel section,
The vowel interval detection result analysis unit, if the audio signal is determined not to be bad speech to row cormorants signal processing unit processing the talker registered with the speech signal or the speech section signal, said input Transmitted voice signal or the above-mentioned voice section signal ,
Illegal voice input determination device. The unauthorized speech input determination device according to claim 1 or 2 ,
A noise suppression and removal unit that suppresses or removes noise in the speech section signal;
The vowel section detection unit detects a vowel section of the speech section signal from which noise is suppressed or removed by the noise suppression and removal unit;
Illegal voice input determination device. Voice section detection unit, and the speech section detection step of generating a speech section signal by extracting a portion of the speech interval from the speech signal input,
A vowel section detecting unit for detecting a vowel section of the speech section signal;
A vowel section detection result analysis unit, including, based on the length of the detected vowel section, determining whether the speech signal is a fraudulent speech, and a vowel section detection result analysis step,
The vowel interval detection result analysis step, when the audio signal is determined not to be bad speech to row cormorants signal processing unit processing the talker recognition using the speech signal or the speech section signal, the Transmitting the input voice signal or the voice interval signal ,
Illegal voice input determination method. Voice section detection unit, and the speech section detection step of generating a speech section signal by extracting a portion of the speech interval from the speech signal input,
A vowel section detecting unit for detecting a vowel section of the speech section signal;
A vowel section detection result analysis unit, including, based on the length of the detected vowel section, determining whether the speech signal is a fraudulent speech, and a vowel section detection result analysis step,
The vowel interval detection result analysis step, when the audio signal is determined not to be bad speech to row cormorants signal processing unit processing the talker registered with the speech signal or the speech section signal, said input Transmitted voice signal or the above-mentioned voice section signal ,
Illegal voice input determination method . Program for causing a computer to function claims 1 as one of the components of the fraud audio input determination equipment 3.

Images